The AI Factory Is Moving: Why Distributed, Energy First Infrastructure Wins

For the last two years, AI infrastructure has been framed as a race for chips, but that framing is now too narrow. The harder problem is increasingly physical: energisation, land, cooling, electrical equipment, and the ability to turn compute demand into live operating capacity. The IEA now projects data centre electricity consumption to reach around 945 TWh by 2030, roughly double 2024 levels, and the U.S. Department of Energy has highlighted connection requests for hyperscale facilities are now in the 300 MW to 1,000 MW range.

The old model is breaking

For most of the cloud era, the industry behaved as though compute picked the location and power would follow, but that assumption is breaking down. AI workloads do not simply add more demand, they change the physical profile of demand. Training concentrates extraordinary amounts of power, cooling, fibre and capital into a small number of sites. Inference then pushes value outwards, closer to users, enterprises and physical systems. AI infrastructure is therefore centralising and decentralising at the same time.

The reason this matters is simple. Power is not constrained in the abstract, it is constrained in specific places, on specific timelines, through specific pieces of infrastructure. The IEA notes that data centres still account for less than 10% of global electricity demand growth to 2030, but because load is highly concentrated, integration becomes much harder locally. Berkeley Lab reports that the average time from interconnection request to commercial operation has risen to more than four years for recently built projects, and DOE says large power transformer lead times are commonly around 36 months and can reach 60 months. In practice, compute is starting to follow power, not the other way around. This is why time to power is increasingly becoming the real commercial clock of AI.

This does not end hyperscale, it changes the architecture around it

None of this means hyperscale stops mattering. Frontier training and large shared inference pools will remain highly concentrated. But once AI moves into live deployment, the question changes. The issue is no longer only how to power the biggest site, it becomes where inference should run when latency, resilience, sovereignty, cost, or continuity of service actually matter. That is the point at which infrastructure stops being a generic capacity exercise and becomes a workload specific architecture decision. This is where edge is increasingly becoming a critical AI strategy.

We can already see this layered model emerging in Europe. The European Commission says 19 AI factories and 13 AI Factory antennas are now operational. At the same time, Deutsche Telekom, Orange, Telefónica, TIM and Vodafone have demonstrated a pan European federated edge continuum, and EURO-3C is deploying more than 70 edge and cloud nodes across more than 13 countries. Europe is not choosing between central AI infrastructure and edge infrastructure, it is building both.

Not every workload belongs at the edge, and pretending otherwise is where the market gets into trouble. Frontier training will stay concentrated. A large share of generic enterprise assistants and batch inference will remain in central pools. Edge wins first where the physical world forces the architecture question, where the cost of distance is operationally real rather than theoretically interesting.

Energy becomes the organising principle

This is why energy is becoming the organising principle for site selection. The right question is no longer just where fibre is dense or where demand is visible today. The right question is where the path to energisation is credible, what combination of grid access, on site generation, storage, and renewables gets capacity live fastest, and how much operational flexibility that site can offer once it is running. DOE now explicitly frames data centre flexibility, on site power generation, storage, and virtual power plants as part of the toolkit for meeting load growth, while the UK Compute Roadmap is exploring new models for the energy infrastructure that powers AI, including renewables, advanced nuclear, and innovative grid solutions.

That is also where storage and orchestration begin to matter. BESS and VPP models are not peripheral to AI infrastructure, they are part of the answer. They create a flexibility layer that can time shift energy, support grid efficiency, and make intermittent or distributed resources more usable for demanding digital loads.

That shift is bigger than site selection alone, it changes what counts as good infrastructure. The most valuable AI sites will not simply be the ones with the largest nameplate capacity. They will be the ones that can secure, shape, store, and optimise energy most intelligently. In the next phase of the market, a megawatt that can be delivered quickly and operated flexibly may be more valuable than a larger theoretical allocation that sits behind years of upgrades.

The AI factory becomes a network

The old model was a campus, the new one is a system. One layer remains concentrated, built for frontier training and large scale shared inference. A second layer becomes regional, positioned where there is a credible energy path, demand density, and commercial logic. A third layer sits closer to physical workflows, in energy networks, mobility corridors, industrial campuses, and other environments where real time decision making has measurable value. The future is not hyperscale or edge. It is a stratified infrastructure model in which each layer serves a distinct technical and economic role.

Energy is likely to be one of the first sectors where this becomes obvious at scale. The UK government has opened a call for evidence on data for AI in the energy system and launched a formal review of AI deployment in electricity networks, covering applications such as planning, balancing, flexibility, optimisation and grid operations. In other words, the grid is not just a constraint on AI growth, it is becoming one of the first major live environments in which distributed AI has to work.

What serious builders do differently

For builders, operators and investors, the implication is clear. Serious AI infrastructure development now starts with a power thesis. A site is no longer credible because it looks well positioned on a map, it is credible if the path to energisation is credible. That means land, substation access, cooling strategy, electrical equipment, permitting, community alignment, and operational readiness all have to be treated as part of one delivery model. At 100MW+ scale, these are no longer secondary considerations, they are the project.

It also means social licence moves closer to the centre of the conversation. As AI campuses become larger, denser, and more visibly tied to local energy systems, infrastructure developers will be judged not only on speed, but on credibility. The winners will be the groups that can build quickly, operate responsibly, and show a clear local and system level benefit.

The real opportunity

The next winners will not be the companies that simply chase the largest cluster in the most familiar market. They will be the ones that can build a network of sites, each aligned to the right energy source, the right latency profile, and the right operating environment. Some of those sites will be massive campuses. Some will be regional AI centres. Some will sit closer to the grid, the factory, the transport corridor, or the enterprise campus.

That is why the next great AI infrastructure advantage will be distributed, energy aware, and workload led. The most valuable AI infrastructure of the next decade may not be the facility that consumes the most power, it may be the system that makes each kilowatt do the most useful work.